Turbocharging Ambient Backscatter Communication

Transcription

1 Turbocharging Ambient Backscatter Communication Aaron N. Parks, Angli Liu, Shyamnath Gollakota, Joshua R. Smith University of Washington {anparks, anglil, gshyam, Co-primary Student Authors ABSTRACT Communication primitives such as coding and multiple antenna processing have provided significant benefits for traditional wireless systems. Existing designs, however, consume significant power and computational resources, and hence cannot be run on low complexity, power constrained backscatter devices. This paper makes two main contributions: (1) we introduce the first multi-antenna cancellation design that operates on backscatter devices while retaining a small form factor and power footprint, (2) we introduce a novel coding mechanism that enables long range communication as well as concurrent transmissions and can be decoded on backscatter devices. We build hardware prototypes of the above designs that can be powered solely using harvested energy from TV and solar sources. The results show that our designs provide benefits for both RFID and ambient backscatter systems: they enable RFID tags to communicate directly with each other at distances of tens of meters and through multiple walls. They also increase the communication rate and range achieved by ambient backscatter systems by 100X and 40X respectively. We believe that this paper represents a substantial leap in the capabilities of backscatter communication. CATEGORIES AND SUBJECT DESCRIPTORS C.2.1 [Network Architecture and Design]: Wireless communication KEYWORDS Backscatter; Internet of Things; Energy harvesting; Wireless 1. INTRODUCTION The last decade has seen significant advances in wireless communication Wi-Fi systems today can achieve bit rates as high as 300 Mbps [2] and cellular networks can operate at ranges of tens of kilometers [3]. These impressive rate and range capabilities have been made possible by communication primitives such as coding mechanisms and multi-antenna processing. These techniques, however, are not suitable for battery-free devices such as RFID tags and ambient power harvesting sensors: Existing multiantenna designs consume power on the order of a few Watts [14]; this is orders of magnitude more power than is available on harvesting devices [24]. Similarly, coding techniques that enhance commu- Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from Permissions@acm.org. SIGCOMM 14, August 17 22, 2014, Chicago, IL, USA. Copyright 2014 ACM /14/08....$ Figure 1 Our Prototype integrates both µmo and µcode in a single design. It can operate using both RFID and TV transmissions. nication range and enable concurrent transmissions (e.g., CDMA) consume substantial power and computational resources that limit their applicability to low-power devices. The key reason for this is that these techniques assume the ability to perform complex digital signal processing multi-antenna processing requires estimating channel parameters and performing operations such as matrix inversion; coding techniques (e.g., CDMA) require performing computationally expensive correlation and synchronization. These methods are infeasible on battery-free devices as obtaining a digital representation of a signal requires powerintensive ADCs, and performing computationally expensive tasks is also expensive in terms of power consumption. In this paper, we ask if it is possible to achieve the benefits of these communication primitives on battery-free devices. The intuition we leverage is as follows: since the power-intensive nature of these techniques is due to their use of ADCs and digital computation, if we instead perform computation in the analog domain we can achieve orders of magnitude reduction in the power consumption. Using this as our guiding principle, we design two communication primitives for backscatter communication: µmo, a multi-antenna receiver that is powered using harvested energy, and µcode, a low-power coding mechanism that enables large communication ranges and concurrent transmissions, and can be decoded using simple analog components on battery-free devices. µmo. To understand our low-power multi-antenna design, consider the ambient backscatter [24] setting in Fig. 2 with two batteryfree devices, Alice and Bob, where Bob communicates with Alice by backscattering the signals from the TV tower. Existing designs consider the TV transmissions to be noise, and hence achieve very low bit rates. Alice, however, can use multiple antennas to separate Bob s reflections from the direct TV signals, thus achieving significantly higher rates. The key challenge however is that conventional multi-antenna decoding hinges on estimating the amplitude and phase of the channel between the transmitters and the receive antennas. Since acquiring phase information is power-intensive, it is not available on backscatter devices [20]. This makes conventional multi-antenna decoding infeasible on backscatter devices. To address this challenge, we present a multi-antenna technique that enables Alice s receiver to decode Bob s backscattered transmission without estimating the channel parameters. The intuition

2 µcode. Coding mechanisms (e.g., CDMA) have been used in conventional radios to significantly increase the range of communication and enable concurrent transmissions. These techniques use pseudorandom codes to encode information at the transmitter; the receiver synchronizes with the transmit signal and decodes the information by correlating with the pseudorandom codes. The challenge in using these designs is that the correlation and synchronization operations are computationally expensive and are difficult to perform on backscatter devices. To address this problem, we introduce a low-power coding scheme for backscatter communication that does not require synchronization. Our intuition is that periodic signals such as sinusoidal waves can be detected without the need for synchronization; thus we design µcode to mimic a sine wave. Specifically, instead of using pseudorandom sequences, we use periodic sequences of alternating zeros and ones as the code. We show that such a design eliminates the need for synchronization operations at the receiver while providing large communication ranges. Further, the receiver can decode using only analog components, thus reducing its power consumption. Finally, we demonstrate that µcode can also enable multiple concurrent transmissions in a network of battery-free devices, without the need for synchronization between transmitters. We build multiple hardware prototypes (one of which is shown in Fig. 1) that integrate both µmo and µcode in a single system. To demonstrate the benefits of our designs for backscatter communication, we build two different prototypes: the first prototype enables ambient backscatter communication using TV transmissions in the 539 MHz frequency range. The second prototype enables RFID tagto-tag communication using RFID reader transmissions in the 915 MHz range. Note that tag-to-tag communication is a novel capability that is in contrast to conventional RFID systems, where the RFID tags communicate with the reader but not with each other. We run experiments in various line-of-sight and non-line-of-sight scenarios at various bit rates from 0.33 bps to 1 Mbps. Our experiments show the following: Figure 2 µmo Decoding. The two-antenna device, Alice, can decode Bob s backscattered information by separating the direct TV transmissions from the backscattered signals. µmo achieves this operation with neither channel estimation nor digital computation. underlying our approach is best explained using an example. Let s(t) be the signal from the TV tower and suppose Bob conveys 1 ( 0 ) bit by reflecting (not reflecting) the TV signals. Alice now receives the following signals on its two antennas: y 1(t) = h rf s(t)+h bbs(t) y 2(t) = h rf s(t)+h bbs(t) where h rf,h rf and h b, h b are the channels from the TV tower and Bob to the two antennas on Alice, respectively. Further, B is 1 when Bob reflects the signals and is 0 otherwise. µmo decodes the backscattered bits without estimating any channel parameters. Specifically, the receiver computes y 1(t) y 2 which is ei- (t) ther h rf h rf or h rf+h b h rf +h b depending on whether the backscattered bit is 0 or 1. Thus, the receiver can decode the backscattered data without performing channel estimation. In 4 we explore the above idea further and describe how the above operation can be performed using simple analog components, without requiring digital computation. Our µmo prototype consumes 422 µw. In comparison, traditional Wi-Fi multi-antenna transceivers consume more than one Watt [14]. µcode, on the other hand, consumes about 8.9 µw. At an ambient illuminance of 330 lux, we could continuously powerµmo andµcode using a photovoltaic array with an area of 8.28 square inches and 0.17 square inches, respectively. By using duty-cycling to reduce average power requirements, the prototypes can also operate on energy from ambient TV signals. µmo increases ambient backscatter s data rate by 100X from 10 kbps to 1 Mbps by providing multi-antenna cancellation; µcode increases ambient backscatter s range by a factor of 40X from two feet to more than 80 feet through coding gain. This significantly expands the range of applications where ambient backscatter can be used. µcode enables RFID tags to communicate with each other at distances of more than 80 feet and between different rooms that are separated by up to three walls. This reduces the need for deploying RFID readers throughout the environment, which is a substantial value proposition. Contributions: We make the following contributions: We present µmo, the first multi-antenna interference cancellation design that can operate on backscatter devices. To do this, we introduce a multi-antenna decoding mechanism that does not require channel estimation and can be achieved using simple analog components. We present µcode, a novel coding mechanism that can significantly increase the backscatter communication range and can be decoded on backscatter devices. Using µcode we enable concurrent transmissions in a network of battery-free devices. Finally, we build hardware prototypes for µmo and µcode and show that they significantly expand the capabilities of backscatter systems. Specifically, we provide orders of magnitude increase in the communication rate and range of ambient backscatter. We also enable RFID tag-to-tag communication at long ranges and in through-the-wall scenarios, which was previously not possible. 2. RELATED WORK Our work is related to prior work in both conventional radio communication and backscatter communication: (a) Radio Communication: There has been significant interest in multi-antenna systems and coding mechanisms to improve the performance of traditional Wi-Fi and cellular systems [36, 12, 40]. These techniques, however, require computationally expensive digital signal processing and power-consuming analog components (e.g., ADCs). In contrast, our work eliminates the need for ADCs and digital signal processing, and achieves multi-antenna cancellation and coding gains on backscatter devices. Our work is also related to low-power MIMO radio designs [38, 8, 34, 18, 22]. These designs increase the energy efficiency of MIMO by optimizing the decoding algorithms [8, 34] and the processor designs [18, 22]. They operate in the digital domain, and thus require power-consuming ADCs. This makes them inapplicable to backscatter devices. Prior work on low-power spread spectrum transceivers [25, 19, 26, 9, 11] has attempted to reduce the power consumption of digital operations such as synchronization and correlation. These systems

3 use traditional radio architectures with power-consuming RF oscillators and ADCs that are not available on a backscatter device. In contrast, we are the first to design and implement CDMA-like codes on backscatter devices that do not require digital signal processing, and hence consume orders of magnitude less power. Finally, our work is related to work in the area of analog signal processing [30, 32, 33]. Before the advent of digital processors, wireless communication was performed on analog circuits. Ultrawide band communication and radar systems still use analog correlators to reduce the cost of hardware components [30, 31, 21]. We build on this foundational work but differ from it in that we are the first to apply this principle to achieve multi-antenna cancellation and channel coding benefits on backscatter devices. Further, we introduce algorithms, techniques, and circuits to achieve this goal. (b) Backscatter Systems: There has been recent interest in improving the performance of backscatter communication [41, 6]. Recent work on ambient backscatter [24] enabled devices to communicate by backscattering existing RF signals like TV transmissions. However, current designs are limited to communication ranges of two feet and communication rates of 10 kbps. Similarly, work on RFID tag-to-tag communication [28] enabled two RFID tags to communicate with each other in the presence of an RFID reader at distances of up to 25 mm. Our work significantly improves the capabilities of these systems. Specifically, using µmo and µcode, we demonstrate ambient backscatter communication at distances greater than 80 feet and communication rates of 1 Mbps using TV signals. We also show the feasibility of RFID tag-to-tag communication at distances of tens of meters, even when multiple walls separate the tags. Recent work on MIMO RFID systems [15, 35, 23, 29, 4, 10] leverage multi-antenna capabilities to improve the throughput and reliability of RFID communication. These systems focus on the communication between a multi-antenna reader and a singleantenna backscatter device, and thus the computationally intensive and power-consuming MIMO operations are performed on the powered reader. Recent information theoretic work [13, 16, 17, 7] characterizes the diversity gains achieved by using multiple antennas on backscatter devices. In contrast, we introduce µmo, a novel multi-antenna technique that can decode the backscattered information by cancelling the interference from ambient signals. We believe that µmo is the first design and implementation of multi-antenna cancellation on a backscatter device. Finally, prior work has proposed to use CDMA to address the problem of collisions in RFID networks [37, 5, 27, 39]. In these systems a code is assigned to each backscatter device, and the powered RFID reader correlates with the code to decode the transmitted data even in the face of collisions. These systems are however designed for uplink communication from the backscatter device to an RFID reader, and do not enable decoding on a low power backscattering device. In contrast, we present µcode, the first coding mechanism that provides the benefits of CDMA and can be decoded on a backscattering device; we also show that µcode can enable concurrent transmissions in a network of battery-free devices. 3. OUR DESIGN PRINCIPLE Our goal is to design primitives that enhance wireless communication capabilities on battery-free devices. The key challenge in achieving this is that wireless primitives such as multi-antenna and coding use digital signal processing that requires power-intensive ADCs and computationally expensive digital operations. To address this challenge, we use the following guiding principle: eliminate the need for digital processing by enabling computation in the analog domain. Such an approach, as we show in the rest of the paper, can provide an order of magnitude reduction in the power consumption of these communication primitives. In the rest of this paper, we describe µmo, our low-power multiantenna backscatter receiver and show how it can improve the communication rate of ambient backscatter. We then describe µcode, our low-power coding technique that is achieved using simple analog components, and show how it can be used to increase the communication range of backscatter systems. Finally, we show that our designs can be used to enable concurrent transmissions in a network of battery-free devices, without the need for synchronization. 4. µmo We first motivate the need for multi-antenna cancellation in ambient backscatter systems and then describe the challenges in achieving µmo. We then describe our receiver design that addresses these challenges. 4.1 Motivation and Challenges Multi-antenna designs are known to provide communication rate gains that are linear with the number of antennas at the receiver; linear improvements, however, might not be a compelling motivation for low-power designs. So we demonstrate next that multiantenna cancellation can provide orders of magnitude communication rate improvements for ambient backscatter systems [24]. Ambient backscatter enables two battery-free devices, Alice and Bob, to communicate with each other by backscattering ambient signals in the environment (e.g., TV transmissions). This is in contrast to traditional backscatter systems (e.g., RFID), where the communication is between a battery-free tag and a powered RFID reader and works only in the presence of an RFID reader. At a high level, the backscattering device encodes a 1 bit by reflecting the incident ambient signals and a 0 bit by absorbing them. The challenge in achieving this is that the ambient signals (e.g., TV) carry information and hence change significantly with time. Prior designs [24] average out the fast-changing TV transmissions to decode the backscattered signals. Averaging removes the variations inherent to the ambient signals, thus allowing the receiver to decode the backscattered signals. The key issue with averaging, however, is that the backscattering device can transmit information only at a rate that is lower than that of the averaging. Specifically, if the receiver averages the TV transmissions over 100 ms, then the backscattering device can only transmit at a maximum rate of 10 bps. Similarly, if the receiver averages transmissions over one millisecond, then the maximum backscatter rate is 1 kbps. Since averaging must be done over a significant period of time to smooth out the variation in the TV transmissions, averaging does not allow rates much higher than 10 kbps [24]. Intuitively, this is because existing designs decode the backscattered information by considering the ambient signals (e.g., TV) as noise. Since the strength of the TV signal is significantly higher than that of the backscattering signal, the channel has a very low SNR and hence achieves a very low communication rate. Using multiple antennas, one can eliminate interference from TV signals, and hence provide orders of magnitude higher bit rates. 1 To see this, consider the 2-antenna receiver in Fig. 2. Suppose s(t) is the signal transmitted by an ambient RF source (e.g., TV tower) and that the backscattering device conveys a one bit by reflecting the TV signals and a zero bit by not reflecting them. Now, the receiver 1 Note that we aim to use multiple antennas for interference cancellation rather than leverage them for antenna diversity. This is because while antenna diversity can reduce the amount of averaging required (e.g., halve it with a two antenna system) it does not eliminate the interference from TV signals.

4 receives the following signals on its two antennas. y 1(t) = h rf s(t)+h bb(t)s(t) y 2(t) = h rf s(t)+h bb(t)s(t) where h rf, h rf and h b, h b are the channels from the RF source (TV tower) and the backscattering device respectively. In principle, if the receiver can estimate the channel parameters, h rf, h rf and h b, h b, it would have two equations in two unknowns, s(t) and B(t); thus it can decode the backscattered information, B(t) by performing a matrix inversion operation on the received signals. Achieving the above operations on a backscatter device, however, is challenging for two main reasons: Channel estimation requires computing both the amplitude and phase on the receiver. Estimating the phase information requires the use of a local oscillator; an oscillator must operate at the carrier frequency, therefore consuming more power than is available. Thus backscatter devices do not have the phase information required to perform channel estimation. Decoding the backscattered signal is done by performing matrix inversion on a digital representation of the received signal. Digitizing the analog signals requires an 8-12 bit resolution ADC that operates at the communication rate; this again consumes more power than is available on a backscatter device. 4.2 µmo Decoding Algorithm µmo can decode backscatter signals with neither phase information nor ADCs. In this section, we describe our algorithm to decode the backscatter signal without estimating the channel parameters. In the next section, we show how the algorithm can be achieved using only analog computation. Our intuition is as follows: Since backscattering creates an additional reflection to the receiver, the effect of backscatter is to change the channel at the receiver. Further, since the reflected signal propagates by different distances to the two antennas at the receiver, the channel changes created on the two antennas are different. Thus, by comparing the relative channel changes across the two antennas, we can identify the bits transmitted by the backscattering device. More formally, µmo first computes the amplitude of the received signal on the two antennas to get: y 1(t) = h rf s(t)+h bb(t)s(t) y 2(t) = h rf s(t)+h bb(t)s(t) where a represents the magnitude of a. Dividing the above equations, we have the following fraction: y 1(t) hrf + hbb(t) = y 2(t) h rf + h b B(t) Note that the above fraction is independent of the TV signal s(t), and hence is not affected by the envelope fluctuations inherent in TV transmissions. Now, since the backscattered bits B(t) are either 1 or 0, the above fraction gives us either h rf. Thus, or h rf+h b h rf h rf +h b the above fraction computation results in two different levels depending on whether the backscattered bit is a zero or a one. To see the above computation in practice, we use a backscatter device to transmit an alternating sequence of zeros and ones at a rate of 1 kbps, by backscattering 539 MHz TV transmissions. We use a USRP equipped with two antennas separated by half a wavelength to capture the received digital signals from a distance of 3 feet. Fig. 3 (a) and (b) plot the received signal on each of the receive antennas. The figures show that it is difficult to see the backscattered signal in the presence of the TV transmissions. Fig. 3 (c) plots the result of dividing the received signals on the two antennas. The plot shows that now the alternating sequence of backscattered bits is visible, allowing the receiver to decode. We note the following key points: The above computation does not guarantee that the output corresponding to a 1 bit is always greater than that corresponding to a 0 bit. In particular, the ratio of the amplitudes on the two antennas depends on h rf, h rf, h b, and h b, which change with device positions. To address this ambiguity, the receiver uses the preamble bits sent by the transmitter to extract the mapping. We note that while noise increases the BER at the receiver, it does not affect the correctness of the above equations. Further, in the above discussion, we assume that the channel is determined by a single channel parameter; since backscattering devices do not use oscillators, they are not frequency selective in their behavior. Thus it is reasonable to approximate the channel at the backscatter device as a single parameter. 4.3 µmo s Hardware Design Next, we describe how µmo achieves the above algorithm using only simple analog components, thus eliminating the need for power-consuming components such as ADCs. The µmo receiver has three main stages: 1) an envelope detection circuit that computes the amplitude of the received signal on each antenna, 2) an analog circuit that imitates the division operation to find the ratio of the signal envelopes received by the two antennas, and 3) a thresholding circuit that extracts the backscattered bits. We explain each of these components in detail. 1) Envelope Detection Circuit. The role of the envelope detector is to track the envelope of the received signal, while eliminating the carrier frequency (539 MHz in case of TV transmissions). The circuit in the figure uses passive components such as resistors, capacitors, and diodes and hence is ultra-low power in nature. The operating principle behind an envelope detector is similar to that used in an RFID tag: diodes act as switches that allow current to flow in the forward direction but not the reverse, capacitors are charge-storage elements and resistors regulate current flow. By choosing the appropriate values for these analog components, we can remove the carrier frequency (e.g., 539 MHz) and leave only the amplitude of the signal. As shown in Fig. 4, µmo performs the envelope detection computation on both the receiver antennas. 2) Divider Circuit. The next step is to divide the amplitudes received on the two antennas. Performing division in a low-power manner is not straightforward, particularly since analog dividers are not a commonly implemented primitive on low power devices. µmo employs simple logarithmic identities to split the division operation into several easily implemented steps, using the following mathematical trick: a b = elog( a b ) = e (log(a) log(b)) Specifically, to compute the quotient of two numbers a and b we can transform them into the logarithmic domain, subtract the two, and then exponentiate to get the result of the computation. µmo imitates the above math in the analog domain to perform division. In particular, µmo uses a log-amplifier on each of the receive signal envelopes to transform the signal amplitudes into the logarithmic domain. At a high level, log-amplifiers make use of the non-linear current-voltage (IV) relationship of a diode to convert a linear domain signal to a logarithmic representation. We then use an analog subtractor to compute the difference between the log-scale signal amplitudes on the two antennas. To get a true linear ratio signal, this difference would need to be transformed back into the linear

5 Power Sample # (a) Signal on the First Antenna Power Sample # (b) Signal on the Second Antenna Power Sample # (c) Result of µmo Division Figure 3 µmo on actual signals. It is difficult to see the backscattered bits on the received signals at the two receiver antennas. Performing the division between the received signals on the two antennas, reveals the backscattered bits. domain using an exponential amplifier (which would also make use of diode properties). However, because our next step is to perform a simple thresholding operation on the signal, we can make use of the fact that the log operation is monotonic and omit the exponential amplifier. Omitting this amplifier saves power by reducing the active component count, and does not impact the performance of the system as the simple thresholding operation is not impacted by the use of log domain signals. 3) Thresholding Circuit. The goal of this circuit is to distinguish between the two voltage levels at the output of the divider circuit. At a high level, the threshold circuit computes the mean of the two voltage levels, and outputs a one bit when the input is greater than the threshold and a zero bit otherwise. To compute this mean value, µmo uses an envelope detector as before. We use specific values for the resistors and the capacitors to compute the mean of the signal envelope. We then use a low-power comparator that outputs a one bit when the divider output is greater than the mean and a zero bit otherwise. As described in 4.2, the receiver uses a known preamble in the packet to flip the bits, if necessary, before decoding. We note that the power consumption of analog components such as log-amplifiers and comparators depends on the bandwidth at which we would like our analog circuit to operate. The bandwidth of the analog circuit is defined as the maximum frequency that the circuit can process without severely attenuating the signals. In our case, we can set the analog bandwidth by selecting low-power amplifier and comparator components that can support a certain bandwidth. In our implementation, we build two different prototypes, one that operates at an analog bandwidth of 1 MHz and the other at an analog bandwidth of 1 khz. 4.4 Putting Things Together The discussion so far focuses on the physical layer of the µmo receiver. The backscatter transmitter uses the same physical layer as RFID systems to transmit messages with preambles and CRC checks. The transmitter can vary the transmitted bit duration and therefore the data rate. The receiver uses the low-power hardware design described above to decode messages, which are then validated by checking the CRC. The transmitter and the receiver use the same link layer mechanisms (including modulation and bit encoding) as prior backscatter systems. To conserve power, µmo s transmit hardware sends an alternating bit sequence long enough (8 bits in our implementation) to wake up the receiver s digital logic system, which then detects bit boundaries and performs framing. 5. µcode We first motivate the need for low-power coding designs for backscatter systems and then describe the challenges in achieving this goal. We then describe µcode s encoding and decoding algorithms that address these challenges. Finally, we present our decoder hardware design that uses only analog components. 5.1 Motivation and Challenges RFID systems have seen a limited adoption, mainly due to limited read range and the resulting requirement for a dense deployment of RFID readers to communicate with RFID tags. If we can enable the RFID tags to communicate with each other at long ranges, we can significantly extend the effective read range by creating a multi-hop network across the RFID tags. Another backscatter technology that is currently limited by range is ambient backscatter [24], which eliminates the need for RFID readers by backscattering ubiquitous ambient signals like TV transmissions. While ambient backscatter does achieve tag-to-tag communication, the communication range of existing work is limited to two feet, which significantly limits its applicability. Thus, increasing the range of backscatter communication is essential to improving the viability of these systems. Coding mechanisms such as CDMA can in principle be used to increase the communication range of backscatter systems. These techniques use a string of pseudorandom bits called chips to encode information bits. For example, the transmitter could represent a zero bit by the chip sequence and a one bit by the chip sequence The 1s and 0s in the above sequences are the chips used by the transmitter to encode the information. To decode these information bits, the receiver correlates the received signal with the chip sequence patterns and converts correlation peaks to data bits. To increase the communication range, the transmitter-receiver pair uses longer chip sequences that effectively increase the signal-to-noise ratio and hence enable long-range communication. Decoding such transmissions on a backscatter device is challenging for two main reasons: The correlation operation required at the receiver is both computationally expensive and requires power-consuming ADCs that are not available on a backscatter device. More important, performing synchronization at the receiver is expensive. Synchronization requires correlating the received signal with a chip sequence for every offset of the received signal. This is not only computationally expensive but also consumes much more power than is available on our target device. So any code that is to be decoded on a backscatter device should satisfy two main properties: 1) It should not require an ADC or digital signal processing operations, and 2) It should work without the need for synchronization. 5.2 µcode s Encoding and Decoding Algorithms µcode can achieve coding gains on the backscatter device without requiring synchronization operations at the receiver. To understand µcode, we first explain the intuition using a simple sinusoidal wave; we then design the encoding and decoding algorithms to operate on the backscattered bits.

6 Figure 4 µmo hardware design. µmo performs the required computation using analog components. Specifically, µmo s hardware has three main components: an envelope detector that removes the carrier frequency (e.g., 539 MHz), an analog division circuit that divides the received signal on the two antennas, and finally a thresholding circuit that outputs the 0 and 1 bits. Intuitively, instead of using a pseudorandom chip sequence, µcode uses a periodic signal to represent the information. To see why this works, consider a periodic sinusoidal signal transmitted at a specific frequency. If the receiver knows the frequency of the sinusoidal signal, it can detect the signal without any phase synchronization with the transmitter. Specifically, suppose the transmitter sends a sine wave, sin(ft+φ), with a frequency f and a phase offset φ. One can easily detect the above sine wave without phase synchronization by performing a dot product operation with the sine and cosine basis functions at the same frequency to compute the In-phase (I) and Quadrature phase (Q) components: I = Q = T sin(ft+φ)sin(ft) t = 1 2 sin(φ) t=0 T sin(ft+φ)cos(ft) t = 1 2 cos(φ) t=0 where T is the duration over which the above dot product operation is computed. 2 Note that from the above equation, I 2 + Q 2 = 1 2 Since Eq. 1 is independent of the phase offsetφthe receiver can detect the transmitted signal without the need for phase synchronization. Further, the receiver can increase its detection sensitivity by increasing the transmit signal duration, T. Thus, a simple periodic sine wave can be detected without requiring phase synchronization, and also provides a form of coding gain to increase reliability. µcode s design builds on the above intuition. However, since our backscatter transmitter has only two states and therefore cannot transmit sine waves, it instead uses a periodic alternating sequence of zeros and ones. Specifically, it encodes the one bit by the chip sequence and the zero bit by the chip sequence Our decoder imitates the I and the Q computation above to detect the alternating one-zero sequence, without the need for synchronization. Specifically, it computes the in-phase component (I) by a dot-product operation with the chip sequence and the quadrature-phase component (Q) by a dot-product operation with a 90 degree offset chip sequence. This is similar to using the sine and the cosine function in the previous scenario. Note that a 90 degree offset is effectively a time offset of half a chip duration. To 2 We replace the integration operation that is typically performed with summation, for ease of exposition. (1) detect the presence of the alternating sequence of zeros and ones, the receiver then computes I + Q. This works independent of synchronization due to the following lemma: LEMMA 5.1. If the transmitter continuously sends an alternating sequence of zero and one chips and the receiver computes inphase (I) and quadrature-phase (Q) components over any duration of N chips, then I + Q = N. The above lemma states that I + Q is constant, independent of receiver synchronization (see appendix for the proof). Thus the receiver can detect an alternating sequence of zeros and ones by computing I + Q over any duration of N chips. To summarize: µcode s Encoding Algorithm. The transmitter uses the following chip sequences to represent the zero and one information bits: 1010 N 10 as bit N 00 as bit 0 where the chips are transmitted at a rate of C. µcode s Decoding Algorithm. The receiver computes I + Q over a duration of n chips. If the majority vote of three adjacent computations is greater than a threshold then the receiver outputs a 1, and 3 a 0 otherwise. The threshold is computed by taking the average of the received signal over bits in the preamble at the beginning of the transmitted packet. The transmitter can control both the chip rate C and the chip length N to achieve different bit rates. The chip rate determines the duration of each chip bit; a higher chip rate corresponds to smaller chip durations. Given a chip rate, longer chip lengths result in larger ranges since they increase the signal-to-noise ratio. We also show in 8 that by controlling these parameters, we can create orthogonal codes that do not require any synchronization and can enable multiple concurrent transmissions. 5.3 µcode Hardware Design Next we describe how µcode achieves the above decoding algorithm in the analog domain, without the need for ADCs. Specifically, as shown in Fig. 5, our receiver design has four main components: an envelope detector circuit, I and Q computation blocks, an addition circuit, and a thresholding circuit. The envelope detector circuit is the same as that described in 4 and is used to remove the carrier frequency (e.g., 539 MHz TV transmissions). We implement the addition operation with an analog adder circuit, and the thresholding circuit is similar to that used in 4.

7 Figure 5 µcode hardware design. Our hardware design performs the decoding operations using analog components, without the need for ADCs. The hardware system has three main components: I/Q correlation circuits that compute I and Q, a summation circuit that computes I + Q, and finally a thresholding circuit that outputs 0 and 1 bits. We focus the rest of this discussion on the I and Q computation circuits. Note that, functionally, both operations are similar to a dot product operation, multiplying two signals and accumulating the result over a time period (in our case the chip length N). The only difference between the I and the Q computations is that the chip sequence is offset by 90 degrees for the Q computation. Thus, we focus on the I computation circuit. At a high level, we achieve the dot product operation (which is a multiplyaccumulate operation) using three key components. First, we send the received signal through an analog inverter to generate the inverted version of the input signal. We then use an analog switch that toggles between the inverted and the non-inverted signal every T seconds, where T is the duration of a chip. This effectively achieves the multiplication operation with the chip sequence. Finally, we use an analog integrator to accumulate the output of this multiplication operation and compute I. The reset function in Fig. 5 is used to reset the circuit every chip sequence length duration to repeat the computation for every transmitted bit. 5.4 Enabling multiple transmissions using µcode µcode can also be used to enable concurrent transmissions in the network without the need to deal with collisions or random access. Specifically, we show in the appendix that µcode satisfies the following property: LEMMA 5.2. A alternating zero-one chip sequence at a chip rate C is orthogonal to all chip sequences at a chip rate 2NC and chip length 2N, where N is a positive integer. The orthogonality property holds independent of the synchronization between the codes. The above lemma implies that transmitters with codes that satisfy the above condition do not interfere with each other. Further, one can create codes with rates C, 2C, 4C, 8C, such that every pair of chip sequences satisfy the above condition. By setting a particular value for C (say 100 Hz) and assigning different codes to different transmitter-receiver pairs, µcode can be used to enable concurrent interference-free transmissions. In 8 we evaluate µcode s ability to enable concurrent transmissions. One aspect of using different codes across transmitter-receiver pairs is the question of distributing these codes across the transmitters. Prior work on CDMA systems [27] has proposed a number of techniques including those that use collision-resistant hash functions to achieve this goal. In principle, we can combineµcode s design with these existing solutions; evaluating this, however, is not in the scope of this paper. 6. FURTHER DISCUSSION We discuss the following aspects of our design: (1) Bit Rate Adaptation. µmo and µcode each provide us with multiple bit rates; the transmitter-receiver pair can perform bit rate adaptation over all these rates to optimize the throughput and to account for channel fluctuations. We note that, for both systems, multiple bit rates can be achieved in hardware simply by changing timing parameters in the digital control logic. Our current prototype implementation allows us to pick different rates across µmo and µcode. Designing bit rate adaptation algorithms however is not in the scope of this paper. (2) Interference with TV receivers. While the receiver designs in this paper increase the range and rate of backscatter communication, the backscattering operation at the transmitter is the same as in prior designs. Specifically, we do not change the strength of the backscattered signal at the transmitter over prior designs, and thus create no additional interference at TV receivers. Thus, similar to [24], our system creates interference at the TV receiver only when it is closer than eight inches from the TV antenna. (3) Application to traditional RFID systems. This paper is focused on enabling communication between battery-free devices using backscatter. However, our designs can also be used to increase the downlink rate and range from an RFID reader to a tag. Current RFID reader-tag systems do not operate across different rooms and are limited to communication ranges that are comparatively smaller; our designs can in principle also provide between-room capabilities to traditional RFID reader to tag links. 7. PROTOTYPE IMPLEMENTATION To evaluate the performance of µmo and µcode in practice, we designed and implemented the system on a printed circuit board as shown in Fig. 1. Our prototype integrates both the µmo and µcode designs on a single board. Fig. 6 shows a block diagram of our receiver prototype; it is a fully reconfigurable platform controlled by firmware executed on the low power MSP430 microcontroller. Both the µmo and µcode sections of the analog circuit are powergated, i.e., they can be enabled or disabled easily to save power. Our prototype has the following features:

8 (a) -30 dbm TV Power at Transmitter (b) -20 dbm TV Power at Transmitter (c) -10 dbm TV Power at Transmitter (d) 0 dbm TV Power at Transmitter Figure 7 µmo BER versus Distance. The figure shows the bit error rate (BER) as a function of distance between the backscattering device and the receiver. The different plots correspond to different ambient RF power levels at the transmitter. The black solid line represents the baseline BER from existing ambient backscatter designs. The figure shows that µmo can enable data rates as high as 1 Mbps at distances between 4-7 feet. This translates to a 100X improvement in the data rates. Figure 6 Block diagram of our receiver prototype Table 1 Power consumption of Analog Signal Processing Max. Rate Range Rx Power µcode 1 kbps >80 ft 8.9µW µmo 1 Mbps 4-7 ft 422µW 1) Flexible switching between µmo and µcode. We use an ultra-lowpower configuration switch as shown in the figure to route signals between two modes of operation: µmo and µcode. When in µmo mode, the microcontroller directs the configuration switch to allow the output of the µmo divider to pass to the next decoding step. The µmo divider circuit is powered on, along with the thresholding circuit. When µmo is active, the µcode correlator is bypassed internally by forcing it into a perpetual reset state, which effectively disables any impact from the correlation circuit. When in µcode mode, the switch is configured to route the signal from one of the two detectors to the correlator, allowing either of the two antennas to be selected as the signal source. The µcode correlator is powered on, along with the thresholding circuit. The microcontroller generates the correlator s reference I and Q chip sequences as well as a periodic reset pulse that clears the correlation result after every chip sequence duration. Note that in our experiments, we use only one receive antenna to evaluate µcode. 2) Trading analog bandwidth for lower power consumption. Every analog communication circuit has an constrained bandwidth, that is, there is a maximum baseband signal bandwidth that a receiver is capable of decoding. The analog bandwidth of a receiver is proportional to the maximum data rate that can be achieved using it. In our prototype, the analog bandwidth is determined by the capabilities of the operational amplifiers and other components we use to implement the analog processing. In general, higher bandwidth allows higher data rates but also increases power consumption. We fabricate two prototypes: one that targets high-rate communication and consumes more power, and another that targets lowrate communication and consumes less power. Specifically, our high data rate prototype targets an analog bandwidth of 1 MHz. To achieve this we make use of the power-efficient TSV6390 operational amplifiers from ST Microelectronics in our signal processing circuits. For the final threshold comparison we use the NCS2200 comparator made by ON semiconductors. The low rate prototype, on the other hand, has an analog bandwidth of 1 khz. To achieve this, we use the BU7265 operational amplifier from Rohm Semiconductor and the TS881 comparator from ST Microelectronics. We use our high rate prototype to demonstrate µmo and our low rate prototype for µcode. Table 1 shows the power consumption of their analog signal processing, including all the analog components. The table shows that µmo consumes about 422 µw of power; this is in contrast to traditional Wi-Fi MIMO implementations [14] that use more than one Watt for a two-antenna receiver. The power consumption numbers for µcode are significantly lower at around 8.9 µw. We note that prior ambient backscatter designs use 0.54 µw of power but are significantly limited to less than two feet of communication range and 10 kbps of bit rates. The higher power consumption of our prototype, however, is not an issue since we use duty cycling to reduce the average power of these analog components, as described next. We also note that one can further reduce these power numbers by an order of magnitude by using application-specific integrated circuits (ASICs). 3) Energy Harvesting and Duty Cycling. Our prototype provides the ability to harvest energy from one or both of the following sources:

9 ambient TV energy, and solar power. In the case of an ambient RF energy harvesting supply, our prototype duty-cycles (periodically alternates between active and sleep states) to allow operation from such a low power source. With a typical TV harvesting power of 100 µw, our 1 MHz µmo device can operate 23% of the time, and the 1 khz µcode system can operate 100% of the time. In the case of a solar photovoltaic power supply, we consider the area requirement of a typical-efficiency solar array such that it can provide sufficient power for continuous operation of µmo and µcode circuits. To do this we measure the power output from an 2.75 by 1.38 inch thin film solar array (manufactured by PowerFilm Inc) in an environment with an illuminance of 330 lux. 330 lux is the minimal office and workplace illuminance specified by the Occupational Safety and Health Administration (OSHA) [1]. Given these measurements, µmo requires a 8.28 square inch solar array and µcode requires a 0.17 square inch solar array to continuously operate in our office environment. 8. EVALUATION We first evaluate the performance of µmo in an ambient backscatter scenario. We then evaluate µcode for both ambient backscatter and RFID tag-to-tag communication in both line-ofsight and through-the-wall scenarios. Finally, we evaluate µcode s ability to enable multiple concurrent transmissions. 8.1 Evaluating µmo We evaluate µmo in an ambient backscatter scenario where two battery-free devices communicate by backscattering ambient TV transmissions. The state-of-the-art ambient backscatter system [24] achieves communication rates of up to 10 kbps at distances of less than two feet. As described in 4, our design can overcome the design limitation of [24] and achieve significantly higher ranges. To demonstrate this, we first evaluate the bit error rate (BER) as a function of distance for various transmitter data rates. We then evaluate the effects of antenna separation on the observed BER values BER versus Distance We evaluate the bit error rate (BER) as a function of the distance between the two ambient backscatter devices. Experiments. We perform the experiments using TV transmissions in the 539 MHz range. We use a single-antenna transmitter and our two-antenna receiver placed at different distances from each other. The two antennas at the receiver are separated by half a wavelength. We compute the bit error rate (BER) observed at the receiver as a function of the distance between the transmitter and the receiver. For each distance value, we repeat the experiments at ten different adjacent positions to account for multipath effects. We compute the average BER at each distance by averaging the BER across the ten locations. Since we are interested in demonstrating high rates, we use only our 1 MHz analog bandwidth prototype. We repeat the experiments for each of four different transmission rates: 1Mbps, 100 kbps, and 10 kbps. To quantify the effect of the TV signal strength on the BER, we measure the TV power level at our backscattering transmitter using a spectrum analyzer, and collect results for four different power levels of -30 dbm, -20 dbm, -10 dbm, and 0 dbm. To compare the results with existing systems, we replicate the design in [24] at a bit rate of 10 kbps and compute the BER as a function of distance for different transmit power levels. Results. Figs. 7(a)-(d) give BER as a function of distance between the backscattering device and receiver for each of four TV power levels. Each plot shows the results for three µmo bit rates and the 10 kbps ambient backscatter design. The plots show the following: BER 1e-02 1e-03 1e Antenna separation (fraction of lambda/2) Figure 9 Effect of separation between the antennas. BER is given as a function of the antenna separation. The bit error rate increases as the distance between the transmitter and the receiver increases. In general, the BER increases with the bit rate. This is expected because at larger distances and higher transmission rates, noise starts affecting decodability. However, the BER values we measure for the µmo system are significantly lower than in prior work on ambient backscatter. As the TV power at the transmitter is reduced, the BER values across all the bit rates increase. This is because, at lower TV signal strengths, the backscatter signal power significantly reduces, decreasing the signal-to-noise ratio at the receiver. However, the BER values measured for µmo are less than 10 2 for distances up to 7 feet; this is true for all tested bit rates provided the TV power is greater than -20 dbm. At an operational BER of 10 2,µmo achieves bit rates of 1 Mbps at a distance of 4 ft across the considered ambient power range. Further, when the ambient power level is greater than -20 dbm, µmo achieves 1 Mbps at distances up to 7 ft. In comparison, a 10 kbps ambient backscatter device (our baseline) can only operate at distances less than two feet across all four power levels. We note that the BER for prior single-antenna ambient backscatter designs [24] is close to 50% at 1 Mbps data rates. This is because, as explained in 4.1, prior designs leverage averaging and hence the backscattering device can transmit information only at a rate that is lower than that of averaging. Since averaging must be done over a significant time period to smooth out variations in the TV transmissions, averaging-based designs [24] do not allow rates much higher than 10 kbps. In contrast, since µmo eliminates the variations in the TV transmissions using division, it can achieve a 100X improvement in the communication rate BER Versus Antenna Separation Next, we evaluate the effect of receive antenna separation on µmo s ability to decode backscattered transmissions. To do this, we fix the positions of the transmitter and the receiver at a distance of three feet from each other. We vary the distance between the two antennas at the receiver from half a wavelength (0.9 feet) to zero feet. The transmitter uses a bit rate of 1 Mbps and is placed in locations where the average TV power is around -10 dbm. We then compute the BER as a function of the antenna separation at the receiver. Fig. 9 plots the results of our experiments. The plot shows that as expected the BER increases as the distance between the two antennas reduces. However, the BER does not see a significant increase until the antenna separation is less than 0.2 half-wavelengths (0.18 feet). This is because while in theory a half-wavelength separation is optimal, in practice smaller receive antenna separations also have good spatial diversity due to wireless multi-path. We also note that, even when the antenna separation is zero, the BER is still This is because, even at this minimal separation, the two antennas still experience a small phase difference that is sufficient to eliminate TV signal fluctuation and decode the backscattered signal.

10 Distance (in ft) TV 1kbps (3.33bps) TV 100bps (0.33bps) RFID 1kbps (3.33bps) RFID 100bps (0.33bps) Power (in dbm) Distance (in ft) TV 100kchip/s (333bps) TV 10kchip/s (33.3bps) RFID 100kchip/s (333bps) RFID 10kchip/s (33.3bps) Power (in dbm) (a) With 1 khz prototype (b) With 1 MHz prototype Figure 8 Communication range versus power in line-of-sight scenarios. The plots show the communication range as a function of the power of the RFID and TV signals as seen by the backscattering device. We plot results for both our 1 khz and 1 MHz prototypes. #Rooms (each 20 ft width) TV 1kchip/s TV 100chip/s RFID 1kchip/s RFID 100chip/s #Rooms (each 20 ft width) TV 100kchip/s TV 10kchip/s RFID 100kchip/s RFID 10kchip/s Power (in dbm) Power (in dbm) (a) With 1 khz prototype (b) With 1 MHz prototype Figure 10 Communication range versus power in non-line-of-sight scenarios. The receiver is moved away from the transmitter into different rooms as shown in Fig. 11. The y-axis represents the number of rooms where the BER is less than Figure 11 Non-line-of-sight experiment locations. We locate the transmitter in the middle of an office at point TX and move the receiver through several adjacent offices to determine the number of walls through which the system can continue to communicate. This is performed for both RFID tag-to-tag communication (continuous wave) and ambient backscatter (TV signal) scenarios. 8.2 Evaluating µcode Next, we evaluate the performance of µcode in practice. As described in 5, µcode achieves two main goals. First, it enables long range RFID tag-to-tag communication by backscattering signals from an RFID reader. Second, it enables ambient backscatter communication at ranges much greater than those demonstrated by existing systems. In this section, we evaluate these goals in both line-of-sight and non-line-of-sight scenarios In Line-of-Sight Scenarios Firstly, we evaluate µcode for RFID tag-to-tag communication and ambient backscatter in line-of-sight scenarios. Experiments. We run experiments in a large indoor space with dimensions 120 ft 40 ft 60 ft. We use both the 1 MHz and 1 khz analog bandwidth receiver prototypes that implement µcode. We pick chip rates of 100 chips/s and 1k chips/s on our 1 khz bandwidth prototype and 10 kchips/s and 100 kchips/s on our 1 MHz bandwidth prototype. At every chip rate we use a chip length of 300 chips to represent an information bit. This combination of chip rates and chip lengths translates to bit rates of 0.33 bps, 3.33 bps, 33.3 bps, and 333 bps, respectively. We measure the range achieved by our system at different RF power levels for the RFID reader or TV signal as measured at our transmitter. To evaluate the benefits of RFID tag-to-tag communication in enable connectivity for tags far from the RFID reader, we place the transmitter (receiver) tag close to (far from) the RFID reader. We vary the power values from -30 dbm to 0 dbm by changing the position of the transmitter. For each combination of power level and transmitted bit rate, we measure the communication range between the transmitter and the receiver. To do this, we move the receiver away from the transmitter and note the distance at which the BER first goes below Note that the receiver moves away from the RFID reader in all our experiments. 3 We repeat the above experiments with TV transmissions using our TV backscatter prototypes. Results. Fig. 8 plots communication range as a function of the signal strength of the reader and TV transmissions as measured at the transmitter. Fig. 8(a) and (b) plot the results for our 1 MHz and 1 KHz prototypes, respectively. The plots show the following: On each of our 1 MHz and 1 khz prototypes, the communication range is higher at lower bit rates. This observation holds for both our TV and RFID backscatter prototypes. This is expected because higher bit rates suffer from higher noise that limits the decoding capability at longer distances. Our 1 MHz analog hardware generally achieves higher ranges than our 1 khz prototype. This is expected, as the 1 MHz hardware has a higher bandwidth and hence has less effective noise. The communication range increases with the power of the signal at the backscattering device. Using the 1 khz prototype, RFID tags can communicate with each other at a rate of 3.33 bps at ranges greater than 90 feet given an incident RF power of 0 dbm. At an incident power level of -15 dbm, the communica- 3 As the receiver moves away from the reader, it can receive neither transmissions nor power from the RFID reader. Our receiver can harvest power from solar and ambient TV signals in these scenarios.